Báo cáo y học: "Gene profiling in white blood cells predicts infliximab responsiveness in rheumatoid arthritis" potx

Abstract As indicators of responsiveness to a tumour necrosis factor TNFα blocking agent infliximab are lacking in rheumatoid arthritis, we have used gene profiling in peripheral blood m

Open Access

Vol 8 No 4

Research article

Gene profiling in white blood cells predicts infliximab

responsiveness in rheumatoid arthritis

Thierry Lequerré1,2,3,4, Anne-Christine Gauthier-Jauneau1,2,3, Carine Bansard2,3,

Céline Derambure1,2,3, Martine Hiron2,3,4, Olivier Vittecoq1,2,3,4, Maryvonne Daveau2,3,4,

Othmane Mejjad1, Alain Daragon1, François Tron2,3,4, Xavier Le Loët1,2,3,4 and

Jean-Philippe Salier2,3,4

1 CHU de Rouen, Hơpitaux de Rouen, Service de Rhumatologie, Rouen, F-76000, France

2 Inserm, U519, Rouen, F-76000, France

3 Université Rouen, Faculté de Médecine-Pharmacie, Institut Fédératif de Recherche Multidisciplinaire sur les Peptides, Rouen, F-76000, France

4 Consortium EGERIE, Rouen, Paris, France

Corresponding author: Jean-Philippe Salier, Jean-Philippe.Salier@univ-rouen.fr

Received: 28 Mar 2006 Revisions requested: 16 May 2006 Revisions received: 23 May 2006 Accepted: 8 Jul 2006 Published: 3 Jul 2006

Arthritis Research & Therapy 2006, 8:R105 (doi:10.1186/ar1990)

This article is online at: http://arthritis-research.com/content/8/4/R105

© 2006 Lequerré et al.; licensee BioMed Central Ltd

This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

As indicators of responsiveness to a tumour necrosis factor

(TNF)α blocking agent (infliximab) are lacking in rheumatoid

arthritis, we have used gene profiling in peripheral blood

mononuclear cells to predict a good versus poor response to

infliximab Thirty three patients with very active disease (Disease

Activity Score 28 >5.1) that resisted weekly methotrexate

therapy were given infliximab at baseline, weeks 2 and 6, and

every 8th week thereafter The patients were categorized as

responders if a change of Disease Activity Score 28 = 1.2 was

obtained at 3 months Mononuclear cell RNAs were collected at

baseline and at three months from responders and

non-responders The baseline RNAs were hybridised to a microarray

of 10,000 non-redundant human cDNAs In 6 responders and 7

non-responders, 41 mRNAs identified by microarray analysis

were expressed as a function of the response to treatment and

an unsupervised hierarchical clustering perfectly separated these responders from non-responders The informativeness of

20 of these 41 transcripts, as measured by qRT-PCR, was re-assessed in 20 other patients The combined levels of these 20 transcripts properly classified 16 out of 20 patients in a leave-one-out procedure, with a sensitivity of 90% and a specificity of 70%, whereas a set of only 8 transcripts properly classified 18/

20 patients Trends for changes in various transcript levels at three months tightly correlated with treatment responsiveness and a down-regulation of specific transcript levels was observed

in responders only Our gene profiling obtained by a non-invasive procedure should now be used to predict the likely responders to an infliximab/methotrexate combination

Introduction

Rheumatoid arthritis (RA) is a chronic, auto-immune and

inflammatory polyarthritis that induces joint damage and

disa-bility Tumour necrosis factor (TNF)α plays a key role in the

associated pathological events and has been identified as a

therapeutic target In fact, TNFα blocking agents (TBAs), such

as infliximab, etanercept, and adalimumab, have revolutionized

the therapeutic care of methotrexate-resistent patients

Vari-ous clinical trials with a TBA/methotrexate combination have

shown efficacy in 60% to 80% of such patients [1-3]

TBAs reduce joint inflammation, slow down joint damage and improve physical function [4,5] Still, 20% to 40% of the RA patients given a TBA/methotrexate combination do not respond to this treatment [1-3] Moreover, TBAs may have side effects and are costly [6] and the efficacy of any given TBA in a given patient is unpredictable [7,8] For these rea-sons, predicting responsiveness to a given TBA or other emerging biotherapies (such as inhibitors of the interleukin-1

or interleukin-6 pathways) would be most useful Markers that have proven informative for RA diagnosis or prognosis, such CRP = C-reactive protein; DAS = disease activity score; DMARD = disease-modifying anti-rheumatic drug; PBMC = peripheral blood mononuclear cell; qRT-PCR = real-time, quantitative reverse transcription PCR; RA = rheumatoid arthritis; SAM = significance analysis of microarrays; TBA = TNFα blocking agent; TNF, tumour necrosis factor.

as C-reactive protein (CRP), erythrocyte sedimentation rate,

autoantibodies (for example, rheumatoid factors and

anti-cyclic citrullinated peptide antibodies), metalloproteinases

and bone proteins cannot predict the responsiveness to TBAs

[9]

Because genetic polymorphisms such as HLA-DR haplotypes

have been associated with a variable natural course of RA and

a heterogeneous response to conventional disease-modifying

anti-rheumatic drugs (DMARDs), a few studies have

attempted to identify genetic markers for TBA efficacy and

they have focused on the promoters of several cytokine genes

[10-12] For example, sequence variation in the TNFα gene

promoter has been associated with a variable response to

inf-liximab [11] However, similar conclusions hold true for

etaner-cept as well [13] and, therefore, such genotypings are useless

for selecting the TBA with greatest benefits [14]

Because response to treatment likely depends on

polymor-phisms at multiple loci [15], genome-wide analysis of gene

expression with cDNA arrays has been recently used to

iden-tify markers of responsiveness in the peripheral blood

mono-nuclear cells (PBMCs) However, the number of such studies

is still very limited [16,17] and very few informative genes have been identified [16] Moreover, in all instances, too few patients per study precluded statistically valid conclusions [17] or a confirmatory analysis in another, independent set of patients [16]

Owing to transcriptome analysis in PBMCs from RA patients,

we have now identified a small subset of transcripts whose combined levels allow one to reliably predict the response to

a infliximab/methotrexate combination in methotrexate-resist-ant patients with very active disease

Materials and methods

Patients

A total of 33 patients, fulfilling the American College of Rheu-matology (ACR) criteria for RA [18] and followed in Rouen University Hospital were included in this study The criteria for patient eligibility were: methotrexate treatment; disease activ-ity score 28 (DAS28) = 5.1 [19]; and resistance to at least one DMARD (methotrexate included) Exclusion criteria were: evolving infectious disease; age <18 years; no contraception;

Table 1

Primers used for qualitative RT-PCR

pregnancy; cancer less than 5 years old; cardiac failure (stage

III-IV of the New York Heart Association); and infliximab allergy

This protocol (numbered 2003/007) was approved by the

eth-ics committee of Haute-Normandie (France) and all

partici-pants signed an informed consent at the time of enrolment For

one month or more before the start of this study every patient

was given fixed amounts of a DMARD and nonsteroidal

anti-inflammatory drug (NSAID) and did not receive any

intra-artic-ular steroid injections During this study, every patient was

given the same doses of methotrexate and prednisone as used

before, and was treated with infliximab (Remicade®,

Schering-Plough, Levallois Perret, France) as recommended by the

manufacturer and the French Drug Agency AFSSAPS

(intra-venous 3 mg/kg infliximab at weeks 0, 2, 6, and every 8th week

thereafter) Before each infiximab infusion, DAS28, plasma CRP level, patient's assessment of pain (0 to 100 mm visual analogue scale), duration of morning stiffness, and physical function scored with the French version of the Health Assess-ment Questionnaire for RA [20] were recorded Just before the 4th infusion (that is, at 3 months), the patients were cate-gorized as responders whenever a change of DAS28 = 1.2 was obtained All others were categorized as non-responders

PBMC isolation and mRNA extraction and labelling

The PBMCs were isolated from venous blood by Ficoll-Hypaque centrifugation and total RNAs were extracted by a standard phenol/chloroform procedure, quality controlled on

an Agilent 2100 Bioanalyzer (Agilent Technologies, Palo Alto,

Table 2

Demographic and clinical data of rheumatoid arthritis patients at entry of study

Subset 1 b (n = 6) Subset 2 (n = 10) Subset 1 (n = 7) Subset 2 (n = 10)

a Categorised as indicated in Materials and methods b Transcript levels were measured by microarray analysis for subset 1 or qRT-PCR for subset

2 c Mean ± standard deviation d Maximally tolerated dose in a given patient e Significant difference between subsets 1 and 2 within

non-responders (p < 0.05, Mann and Whitney's non-parametric test) In this table, all other comparisons were non-significant Anti-CCP abs,

anti-cyclic citrullinated peptide antibodies; NSAID, non-steroidal anti-inflammatory drugs; RA, rheumatoid arthritis.

Table 3

Clinical data at baseline and at 3 months

Subset 1 Subset 2 Subset 1 Subset 2

Baseline 3 months a Baseline 3 months a Baseline 3 months a Baseline 3 months a

Morning stiffness

(minutes) b 245 ± 126 c 35 ± 24.5 c 210 ± 81 58 ± 70.2 c 179 ± 159 66.4 ± 86 c 133 ± 84 62 ± 67.6 c

DAS28 b 6.4 ± 1.0 4.2 ± 0.9 c 6.2 ± 0.7 3.8 ± 0.6 c 5.7 ± 0.8 5.3 ± 1.0 5.5 ± 1.0 4.9 ± 1.0 c

Pain (0–100 mm VAS) 59.3 ± 20.3 29.3 ± 9.3 c 62.5 ± 15.5 31.3 ± 14.5 c 69.3 ± 13.1 54.1 ± 22.1 60.9 ± 11.4 40.6 ± 18.4 c

ESR (mm/hour) 44 ± 26.2 27 ± 20.3 c 27.2 ± 15.7 11.3 ± 5.2 c 35.7 ± 25.7 28.3 ± 15.3 24.1 ± 11.5 27.8 ± 19.2 CRP (mg/l) b 42 ± 29.8 20 ± 15.7 c 28.6 ± 19.7 6.2 ± 6.1 c 18.5 ± 12.7 13 ± 8.2 15.8 ± 15.6 11 ± 7.3 HAQ score (0–3 scale) 1.6 ± 0.4 0.9 ± 0.5 c 1.8 ± 0.7 1.2 ± 0.7 c 1.6 ± 0.4 1.2 ± 0.3 1.5 ± 0.4 1.5 ± 0.4

Values are mean ± standard deviation a Response assessed just before the fourth infliximab/methotrexate infusion Significant differences between groups are as follows: bdifference between all responders versus non-responders at baseline (0.03 <p < 0.05, Mann and Whitney's test); other comparisons were non-significant (p = 0.58); cDifference at baseline versus 3 months in this subset (p < 0.05, paired Wilcoxon's test)

CRP, C-reactive protein; DAS, disease activity score; ESR, erythrocyte sedimentation rate; HAQ, health assessment questionnaire; VAS, visual analogue scale (patient's assessment of pain).

USA) and frozen at -80°C until further use An internal,

arbi-trary standard was made of a mixture of total RNAs from

PBMCs taken from three healthy donors The oligodT-primed

poly(A) mRNAs were labelled with [α33P]dCTP as previously

described [21], and the resulting, labelled cDNAs were

imme-diately used for hybridisation

Transcriptome analysis and qRT-PCR

Our array covering 12,000 cDNA probes for 10,000

non-redundant genes and various negative controls as well as

nylon arraying of PCR-amplified probes and hybridisation of

[α33P]dCTP-labelled mRNAs have all been extensively

described and validated in a previous report [21] Briefly,

cDNA probes selected on the basis of a tissue-preferred

expression in liver corresponded to genes with a

liver-restricted expression (10% of the probes) as well as genes

with an hepatic expression along with a broad expression in

some (50%) or many non-hepatic tissues (40%) [21] All

arrays were made from a single batch of cDNA probes Every RNA sample was hybridised at least twice on separate arrays Whenever necessary, the sequence of cDNA probes was con-trolled with an ABI3100 capillary sequencer (Applied Biosys-tems APPLERA-France, Courtaboeuf, France) Real-time, quantitative reverse transcription PCR (qRT-PCR) of mRNAs and normalization with the 18S RNA amount were done in duplicate as described [21] The primers designed with the Primer3 software [22] are listed in Table 1

Image analysis and data mining

Image analysis with the XDotsReader software, version 1.8 (COSE, Le Bourget, France), subtractions of noise and spot background, and image normalization with the median value of all signals per image were done exactly as previously detailed [21] A transcript was considered to be expressed if at least two hybridisations provided a positive signal The resulting, normalized values were used for a selection of significantly

Figure 1

Clustering of rheumatoid arthritis patients as responders versus non-responders

Clustering of rheumatoid arthritis patients as responders versus non-responders Transcripts in peripheral blood mononuclear cells from six respond-ers (R) or seven non-respondrespond-ers (NR) who were included in two training subsets (subset 1 in text and Tables 2 and 3) were studied by microarray

analysis Informative transcripts as selected by a statistical analysis (t test, 25 transcripts; significance analysis of microarrays (SAM), 37 transcripts)

were next used for an unsupervised hierarchical clustering of the same 13 patients, listed as columns The gene names are listed as rows (expressed sequence tags are noted with a plain, five to six digit IMAGE clone number) The genes are underlined whenever they were selected by both SAM

and t test Transcript levels are expressed as ratios (level in sample/level in internal, arbitrary standard) Scale bar (log2 ratio): decreased (green),

increased (red) or identical (black) ratio in sample versus standard (grey squares are missing values).

Table 4

Transcripts as predictors of infliximab responsiveness

IMAGE clone a Encoded protein Symbol b Gene localisation SAM c t testd

77684 CytP450, family 3, subfamily A, polypeptide 4 CYP3A4 7q21.1 -2.90 <10 -4

417137 A kinase (PRKA) anchor protein 9 AKAP9 7q21-q22 -2.83 0.002

415079 Hypothetical protein DKFZp566M1046 - - -2.78 0.001

1848509 RP1 containing part of thyroid hormone

receptor-associated protein 3 THRAP3 1p34.3 -2.56 NS

198699 C-X-C chemokine ligand 5 (ENA78) CXCL5 4q12-q13 -2.50 NS

730048 Ribosomal protein SA (37LRP) LAMR1 3p21.3 -2.43 0.007

56923 F-box protein 5 FBXO5 6q25-q26 -2.42 0.006

1524020 RAS guanyl releasing protein 3 (calcium and

DAG-regulated)

756784 WD repeat domain 39 WDR39 2q11.2 -2.40 NS

124452

6-Phosphofructo-2-kinase/fructose-2,6-biphosphatase 4

724887 Major HLA, class II, DP beta 1 HLA-DPB1 6p21.3 -2.32 <10 -4

416493 Ribosomal protein L35 RPL35 9q34.1 -2.25 NS

191599 Hypothetical protein FLJ13614 - - -2.23 0.006

726045 Ribosomal protein S16 RPS16 19q13.1 -2.24 NS

772993 Similar to 40S ribosomal protein S28 RPS28 19p13.2 -2.23 NS

110169 Proteasome subunit β type 9 (LMP2) PSMB9 6p21.3 -2.17 0.006

346678 Musculoskeletal, embryonic nucleic protein 1 MUSTN1 3p21.1 -2.16 NS

741027 Vinexin β (SH3-containing adaptor molecule-1) SCAM-1 8p21.3 -2.15 NS

428222 EGF receptor pathway substrate 15 EPS15 1p32 -2.12 0.003

740374 Transducin (beta)-like 2 TBL2 7q11.23 -2.12 Ns

774502 Protein tyrosine phosphatase, non-receptor type

12

320298 Membrane-type 1 matrix metalloprotein

cytoplasmic tail binding protein 1

148134 RP1-29K1 containing KiAA0426 - - NS 0.002

127203 CytP450, family 4, subfamily F, polypeptide 12 CYP4F12 19p13.1 NS 0.005

810626 Cytochrome c oxidase subunit VIIa polypeptide 2

like

486624 ELMO domain containing 2 ELMOD2 4q31.21 +1.85 NS

357960 Mitochondrial ribosomal protein L22 MRPL22 5q33.1-q33.3 +1.99 0.009

82303 Hypothetical protein BC009264 - - +2.12 NS

247517 Mucin and cadherin-like MUCDHL 11p15.5 +2.40 NS

194455 Membrane cofactor protein (CD46) MCP 1q32 +2.30 0.005

regulated mRNAs, that is, those with an abundance that

dif-fered in two or more comparisons between two samples,

using a funnel-shaped confidence interval (p < 0.05)

calcu-lated from every mRNA detected per hybridisation [21] This

results in a false discovery rate that is below 10% of the total

number of regulated mRNAs Statistical analyses were done

with the R software [23] The TIGR Multiexperiment viewer

(Tmev version 2.2) [24] was used for unsupervised

hierarchi-cal clustering (HC) using the average dot product and

com-plete linkage options, the leave-one-out cross-validation, and

the supervised statistical tool Significance Analysis of

Micro-arrays (SAM) for identification of discriminant transcripts [25]

with a false discovery rate set at <1% Information about our

clinical and experimental data complies with the

recommenda-tions for the minimum information about microarray

experi-ments (MIAME) and the raw data have been deposited

(accession number GSE3592) in the GEO repository [26]

Results

RA patients and response to treatment

We categorized patients into two groups, responders (R1 to

R16) and non-responders (NR1 to NR17) to an infliximab/

methotrexate combination, at three months according to the

EULAR criteria, as recommended [18] Tables 2 and 3 provide

demographic and clinical information for these 33 patients, at

entry and at 3 months The average disease duration was 11

to 12 years and the DAS28 score indicated that all these

patients had a high level of RA activity, which fits with their

resistance to one or more DMARDs Before treatment, three

variables (morning stiffness, DAS28, CRP level) were slightly

different in responders compared to non-responders

Follow-ing treatment, the DAS28 score significantly improved at 3

months in responders (average decrease 2.3) whereas it

remained high in non-responders (average decrease 0.4)

Patients in both groups were randomly separated into either a

training subset (subset 1) for transcriptome analysis or a

vali-dation subset (subset 2) for qRT-PCR At this stage, we paid

attention to retaining a relatively large number of patients in

subset 2 of both groups As noted in Tables 2 and 3, most

fea-tures did not significantly differ between the paired subsets 1

and 2

Gene profiling in pre-treatment PBMCs correlates with treatment responsiveness

Gene profiling in PBMCs was studied in the two training sub-sets (subset 1) of the responders and non-responders groups (a total of 13 patients) On average, 5,282 ± 1,253 transcripts were detected in PBMCs, with 86% overlap in transcript iden-tities between responders and non-responders (data not shown) To precisely identify the transcripts that were differen-tially regulated in responders compared to non-responders,

we first selected every transcript whose level in at least one responder was significantly different from the median value in non-responders or vice versa This was assessed with a

fun-nel-shaped confidence interval (see Materials and methods; p

< 0.05) and resulted in 2,239 transcripts with an abnormal level in at least 1 out of these 13 patients From these 2,239 transcripts, we next selected every transcript whose variation between responders and non-responders was statistically

sig-nificant according to a t test (25 transcripts) and/or SAM (37

transcripts); these transcripts are listed in Figure 1 (total, 41

transcripts; overlap between t test and SAM selections, 21

transcripts) and detailed in Table 4 The identity of the corre-sponding microarray cDNA probes was verified by sequenc-ing Finally, we performed an unsupervised hierarchical clustering of the 13 patients above (subset 1) This was based

on the levels of the 25 or 37 transcripts indicated above that,

in both instances, resulted in a perfect separation of the responders and non-responders into two major clusters (Fig-ure 1)

We wished to confirm that a combination of the above tran-script levels could be used as a predictor of responsiveness For this purpose, we aimed to measure the levels of the above

41 transcripts by qRT-PCR and compare them between our two validation subsets (subset 2) from the responder and non-responder groups (a total of 20 patients) However, among these 41 transcripts, 12 putative transcripts were identified by only one IMAGE clone without knowledge of the intron/exon structure and, therefore, they were not retained Moreover, among the 29 remaining transcripts, 9 of them failed to pro-vide reliable data by qRT-PCR, despite repeated attempts with various primers Eventually, 20 out of our 41 transcripts could

be reliably quantified by qRT-PCR As shown in Figure 2a, an unsupervised hierarchical clustering of the 20 patients in sub-set 2 from the two groups, as based upon these 20 transcript levels, resulted in two major clusters of responders versus

244896 Aminoadipate aminotransferase AADAT 4q33 +2.52 0.002

a IMAGE clone number as a unique identifier b Bold indicates a transcript that was further tested by qRT-PCR c Significance analysis of microarrays (SAM) value as an indicator of significant transcript variation in responders versus non-responders A positive or negative value indicates an over- or underexpression at baseline in responders versus non-responders, respectively dP value of a t test as an indicator of significant transcript variation

in responders versus non-responders NS, non- significant.

Table 4 (Continued)

Transcripts as predictors of infliximab responsiveness

non-responders, with 5 misclassified patients (NR8, NR12,

NR17, R13, R16) Despite being informative, such a

hierarchi-cal clustering lacks statistihierarchi-cal power, and the efficiency of the

above set of 20 transcripts for patient classification was thus

further evaluated by leave-one-out cross-validation [24] This

procedure identified 4 misclassified patients and indicated

that this set of transcripts provides 90% sensitivity and 70%

specificity for identification of responders and non-responders

(Table 5)

To determine the minimal number of transcripts that should be

measured for an acceptable prediction of responsiveness, we

tested a series of combinations of transcripts in the 20

patients from each subset 2, and we varied the number and

identity of the transcripts actually used (data not shown) With

a given set of only 8 transcripts, 16 out of 20 patients could

be correctly classified as responders or non-responders by

hierarchical clustering (Figure 2b) Finally, leave-one-out

cross-validation (Table 5) identified only two misclassified

patients and indicated that a given set of 8 transcripts as a

pre-dictor of responsiveness was at least as accurate as the set of

20 transcripts above

Post-treatment transcript levels correlate with treatment

responsiveness

We investigated whether the differences in transcript levels

seen in responders compared to non-responders at baseline

were also retained at three months The data obtained by

qRT-PCR with PBMCs are presented in Figure 3 In responders, 18

out of 20 transcripts (90%) exhibited a trend towards an

increased level at 3 months, although the differences with

respect to the levels at baseline were not significant Strikingly,

in non-responders, 19 out of 20 transcripts (95%) exhibited an opposite trend, that is, a decreased level at 3 months, and this difference was statistically significant for each of 8 transcripts (Figure 3) Overall, the differences in numbers of up- versus down-regulated transcripts in responders versus non-responders were highly significant, whether considering only the number of transcripts with a significant difference at

base-line versus 3 months (n = 8, p = 3.10-3, Fisher's exact test) or considering the complete set of transcripts and associated

trends (n = 20, p < 10-4 by Fisher's exact test, or p = 0.007 by

analysis of variance) This argues for a regulation of the corre-sponding genes by one (or more) TNFα-dependent pathway(s)

Discussion

The small set of biological markers usually used for RA diag-nosis or progdiag-nosis is unable to predict individual responsive-ness to TBA [14] Therefore, to enable such a prediction, global approaches based on proteomics or transcriptomics have been recently considered [27,28] However, in the con-text of RA, proteomic analysis is still under development [27] Moreover, very few informative transcripts have been identified

by gene profiling [16] and the few studies that used this approach have relied on the differences in transcript levels measured at baseline versus two to three days after treatment onset [17] This required exposure of every patient to treat-ment Furthermore, the narrow time frame of this procedure may blur some significant but late variations with respect to baseline, which eventually limits transcript informativeness In contrast, we have now measured transcript levels at baseline

Figure 2

Validation of a narrow selection of transcripts as a tool for clustering responders versus non-responders

Validation of a narrow selection of transcripts as a tool for clustering responders versus non-responders Ten responders (R) and ten non-respond-ers (NR) were included in two validation subsets (subset 2 in text and Tables 2 and 3) In any given sample of peripheral blood mononuclear cells, the abundances of informative transcripts were determined by qRT-PCR and normalized with the corresponding 18S RNA level Unsupervised

hier-archical clusterings obtained with (a) 20 or (b) 8 selected transcripts are shown Expression of transcript levels and scale bar are as in Figure 1.

as the single predictor of responsiveness In clinical practice, prediction can then be done without any exposure to treat-ment, which enables it to be restricted to responders Three months of treatment was chosen as the endpoint of our study, as recently recommended by international experts [29], because the objective of an efficient RA treatment is a rapid response Should this early evaluation at three months dis-close a moderate or absent response, this procedure allows another treatment to be used as early as possible Also, using the DAS28 evolution at three months for classifying our 33 patients as responders or non-responders turned out to be quite reliable in the long run Indeed, 22 out of 33 patients could be followed for three more years and their infliximab responsiveness, or lack thereof, did not vary over this period, even when increasing infliximab amount and frequency in non-responders (data not shown)

We aimed to identify a list of transcripts whose combined lev-els could be related to infliximab/methotrexate responsive-ness In fact, infliximab used alone is known to be efficient only for a short durationbecause the rapid production of anti-inflixi-mab antibodies counteracts the drug's effect, whereas meth-otrexate advantageously limits this occurrence The mixture of

a cytokine inhibitor (infliximab) and an inhibitor of cell prolifer-ation (methotrexate) is likely to regulate or even co-regulate a complex set of genes; this is a limitation if an understanding of some underlying events in RA is desired

Gene expression was measured in PBMCs because this is an acknowledged, non-invasive procedure for diagnosis or prog-nosis of autoimmune diseases [30] Specifically, in the context

of RA, PBMCs as a surrogate tissue are advantageous as they allow for screening in any subject, whereas synovium is ame-nable to analysis in only a few patients However, a drawback

of such PBMC analysis is the lack of a clear-cut relationship between PBMCs and the affected synovium, which prevents the resulting data from providing an understanding of the RA-associated events in joints Also, we analysed the PBMC tran-scriptome with an arbitrary collection of approximately 10,000 cDNA probes [21] Since this restrictive procedure cannot measure every transcript expressed in the PBMCs, it does not intend to provide a genome-wide view of the RA-associated gene dysregulations in this tissue Yet, this approach is quite acceptable when inferring prognosis from gene profiling is the major task

Overall, the present study was not designed primarily to increase our understanding of RA physiopathology but is mostly suited to the predictive use of some combined tran-script levels Our data illustrate that a non-invasive transcrip-tome analysis done in PBMCs with an array of probes devoid

of a specific selection towards the disease under study ena-bles the efficient prediction of treatment responsiveness Whether these conclusions are solid whatever the microarray/

Figure 3

Relative transcript levels at baseline versus three months in responders

or non-responders

Relative transcript levels at baseline versus three months in responders

or non-responders The patients and transcripts are as in Figure 2a For

every transcript, the 4 levels (median value) shown at baseline and after

3 months in responders and non-responders are expressed as a

per-centage of the median level at baseline in responders (100%)

Signifi-cant differences are all noted in the non-responder panel: asterisk

outside closed bar, difference in non-responders at baseline versus

three months (p < 0.05, paired Wilcoxon's test); asterisk within open

bar, difference at baseline in responders versus non-responders (p <

0.05, Mann and Whitney's test); asterisk within closed bar, difference

at 3 months in responders versus non-responders (p < 0.05, Mann and

Whitney's test) In any patient group, a trend towards an increased or

decreased level was considered whenever the value at 3 months was,

respectively, above or below the value at baseline, whatever the

differ-ence of these values Note that standard deviations are not shown

because they are useless for non-parametric statistical tests.

qRT-PCR platforms used, depend on a restricted PBMC

sub-population, or, above all, are useful in the context of an actual

therapeutic decision, remains to be tested

By t test and/or SAM, we identified a short list of 25 to 37

tran-scripts whose combined expression levels in PBMCs are an

efficient discriminator of responders versus non-responders to

infliximab/methotrexate Many of the 25 transcripts identified

by t test were no longer significant when using Bonferroni's

correction to adjust statistics for the multiple transcripts

ana-lysed, but Bonferroni's correction has been recognized as a

drastic one when used in this context, which contrasts with the

SAM-associated false discovery rate [31] Moreover, the t test

and SAM cross-validated each other for most of the 20

tran-scripts eventually selected for qRT-PCR as 13 out of 20

(65%) such transcripts were significant with both tests (Table

4) Measuring these 20 transcript levels by qRT-PCR

indi-cated that their performance as a predictor of responsiveness

was equal to that obtained with 37 transcripts Ultimately, a

given combination of 8 selected transcripts (75% of them

being significant by t test and SAM) as a predictor of

respon-siveness was as powerful as any higher number of transcripts

This observation that a given combination of very few

tran-scripts can equal or even outperform the predictive strength of

a higher number of transcripts has also been reported in

another context, namely the response to hepatitis C treatment

[32] This small size for an informative gene set is most

encour-aging when the need comes for the development of a reliable,

fast and cheap assay for measuring informative transcript

lev-els in a clinical setting

Consistent with the limitations noted above, our list of 29

tran-scripts did not disclose any significant series of trantran-scripts

whose altered levels could point to the physiopathological

importance of a predominating function or pathway Indeed,

these transcripts covered such diverse proteins and functions as: ribosomal components (LAMR1, MRPL22, RPL35, RPS16, RPS28), which may suggest the existence of a TNFα-dependent pathway in the control of translation; cell adhesion and inhibition of cell migration/invasion (LAMR1, MUCDHL, MTCPB1); cytochromes (CYP3A4, CYP4F12) and cyto-chrome oxidase (COX7A2L); proteasome-mediated proteoly-sis (FBXO5, PSMB9); various enzymes (AADAT, PFKFB4); intra- or extracellular signalling (AKAP9, CXCL5, PTPN12, RASGRP3, TBL2, THRAP3), including regulators of the ERK pathway (EPS15, SCAM-1); and innate or adaptive immunity (KNG1, MCP, PSMB9, HLA-DPB1) Two transcripts, namely MUSTN1 and HLA-DPB1, are noteworthy; the MUSTN1 tran-script codes for a protein involved in bone development and

regeneration [33] and some alleles of the HLA-DPB1 gene

have been associated with a relatively high risk of RA occur-rence [34]

The opposite variations in transcript levels seen in responders compared to non-responders at three months strongly sug-gest that the informative transcripts retained in our study orig-inated from TNFα-regulated genes In fact, TNFα-dependent

expression of the CXCL5, CYP3A4, LAMR1, MCP, and PSMB9 genes, as noted here, has been previously described

[35-40] However, only two of our transcripts, namely MCP and PTPN12, are found among lists of genes that are directly regulated by the TNFα/NFκB pathway, whether in RA [41] or

in another context [42,43] Therefore, it is likely that most of our transcripts are indirect TNFα targets This view fits with the fact that the opposite variations in responders versus non-responders were observed weeks after the start of TBA The reason why the transcript levels exhibited a limited trend to up-regulation at three months in responders along with a predom-inating repression in non-responders (Figure 3) also fits with indirect TNFα target genes, whose regulation would depend

Table 5

Performance of the number of transcripts for prediction of responsiveness

Number of selected transcripts a

a As listed in Figure 2a, b b By leave-one-out cross-validation with 20 patients, including 10 non-responders (NR) and 10 responders (R) (referred

to as validation subset 2 in the text) cP < 0.05 indicates a significant link between transcript-based classification (R versus NR) and actual

responsiveness.

on one or more TNFα-dependent transcriptional repressor(s).

The difference in responders versus non-responders could

then result from genetic polymorphisms in binding sites for

such repressors

This situation of variations in binding of transcription factors

has been previously described in RA [11,44] Notably, the

-308G/G genotype of the TNFα gene promoter is known to be

associated with a better response to infliximab compared to

the -308A/G or A/A genotype [11] Other binding sites for

repressor(s) could be located in any gene that belongs to the

pathway from TNFα signalling to its indirect target genes

whose transcripts were found here If so, identifying such

bind-ing site polymorphisms that could predict the extent of

respon-siveness to TBA deserves further studies Beyond this, it might

well be worth combining the HLA-DRB1 genotype, itself a

predictor of responsiveness to methotrexate/sulphasalazine/

hydroxychloroquine in RA [45], with our measure of

informa-tive transcript levels, as this might enhance the predicinforma-tive

power of such indicators

Conclusion

The combined levels of a small set of discriminative transcripts

have provided for the first time a tool for the prediction of

inf-liximab/methotrexate efficacy in patients with long standing

(11 to 12 years) and very active RA It remains to be seen

whether our predictive approach can prove useful in patients

with recent and/or moderate RA activity or in non-responders

given a higher dose of infliximab (>3 mg/kg) Other future

studies should identify further gene profiles whose changes

correlate with a responsiveness to other TBAs or treatments,

such as interleukin-1 receptor antagonists [46] Ultimately, we

anticipate that a small series of parallel tests for such

drug-specific combinations of transcripts, as quantified on a

specif-ically designed DNA chip, should allow one to select the most

appropriate treatment for every RA patient, with the resulting

and beneficial eradication of the non-responder or moderate

responder phenotypes

Competing interests

A patent application EP 06290789.4 for the set of 20 or 8

transcripts with predictive power (Figure 2) has been

depos-ited by Inserm The authors declare that they have no

compet-ing interests

Authors' contributions

TL, XLL and JPS were responsible for designing the study and

writing the manuscript OV, OM, AD, and XLL were

responsi-ble for clinical coordination, access to samples, RA evaluation

and manuscript improvements ACGJ, CB, CD, and MH were

responsible for microarrays, qRT-PCRs and data analysis MD

and FT provided numerous manuscript improvements

Acknowledgements

CB is the recipient of a fellowship from the French Ministry for Research This work was supported in part by a grant from Conseil Régional de Haute-Normandie (France).

References

1 Bathon JM, Martin RW, Fleischmann RM, Tesser JR, Schiff MH, Keystone EC, Genovese MC, Wasko MC, Moreland LW, Weaver

AL, et al.: A comparison of etanercept and methotrexate in patients with early rheumatoid arthritis N Engl J Med 2000,

343:1586-1593.

2 Keystone EC, Kavanaugh AF, Sharp JT, Tannenbaum H, Hua Y,

Teoh LS, Fischkoff SA, Chartash EK: Radiographic, clinical, and functional outcomes of treatment with adalimumab (a human anti-tumour necrosis factor monoclonal antibody) in patients with active rheumatoid arthritis receiving concomitant meth-otrexate therapy: a randomized, placebo-controlled, 52-week

trial Arthritis Rheum 2004, 50:1400-1411.

3 Maini RN, Breedveld FC, Kalden JR, Smolen JS, Furst D, Weisman

MH, St Clair EW, Keenan GF, van der Heijde D, Marsters PA,

Lip-sky PE: Anti-tumour necrosis factor trial in rheumatoid arthritis with concomitant therapy study group Sustained improve-ment over two years in physical function, structural damage, and signs and symptoms among patients with rheumatoid

arthritis treated with infliximab and methotrexate Arthritis Rheum 2004, 50:1051-1065.

4 Klareskog L, van der Heijde D, de Jager JP, Gough A, Kalden J,

Malaise M, Martin Mola E, Pavelka K, Sany J, Settas L, et al.:

Ther-apeutic effect of the combination of etanercept and meth-otrexate compared with each treatment alone in patients with rheumatoid arthritis: double-blind randomised controlled trial.

Lancet 2004, 363:675-681.

5 Smolen JS, Han C, Bala M, Maini RN, Kalden JR, van der Heijde D,

Breedveld FC, Furst DE, Lipsky PE, ATTRACT Study Group: Evi-dence of radiographic benefit of treatment with infliximab plus methotrexate in rheumatoid arthritis patients who had no clin-ical improvement: a detailed subanalysis of data from the anti-tumour necrosis factor trial in rheumatoid arthritis with

con-comitant therapy study Arthritis Rheum 2005, 52:1020-1030.

6 den Broeder A, van de Putte L, Rau R, Schattenkirchner M, Van Riel P, Sander O, Binder C, Fenner H, Bankmann Y, Velagapudi R,

et al.: A single dose, placebo controlled study of the fully

human anti-tumour necrosis factor-alpha antibody

adalimu-mab (D2E7) in patients with rheumatoid arthritis J Rheumatol

2002, 29:2288-2298.

7. Ang HT, Helfgott S: Do the clinical responses and complica-tions following etanercept or infliximab therapy predict similar outcomes with the other tumour necrosis factor-alpha

antag-onists in patients with rheumatoid arthritis? J Rheumatol

2003, 30:2315-2318.

8 Haraoui B, Keystone EC, Thorne JC, Pope JE, Chen I, Asare CG,

Leff JA: Clinical outcomes of patients with rheumatoid arthritis

after switching from infliximab to etanercept J Rheumatol

2004, 31:2356-2359.

9 Lequerré T, Jouen F, Brazier M, Clayssens S, Klemmer N, Ménard

J-F, Mejjad O, Daragon A, Tron F, Le Loët X, Vittecoq O: Autoan-tibodies, metalloproteinases and bone markers in rheumatoid arthritis patients are unable to predict their responses to

infliximab Rheumatology 2006 in press.

10 Kang CP, Lee KW, Yoo DH, Kang C, Bae SC: The influence of a polymorphism at position -857 of the tumour necrosis factor alpha gene on clinical response to etanercept therapy in

rheu-matoid arthritis Rheumatology 2005, 44:547-552.

11 Mugnier B, Balandraud N, Darque A, Roudier C, Roudier J, Reviron

D: Polymorphism at position -308 of the tumour necrosis fac-tor alpha gene influences outcome of infliximab therapy in

rheumatoid arthritis Arthritis Rheum 2003, 48:1849-1852.

12 Cuchacovich M, Ferreira L, Aliste M, Soto L, Cuenca J, Cruzat A,

Gatica H, Schiattino I, Perez C, Aguirre A, et al.: Tumour necrosis

factor-alpha (alpha) levels and influence of -308 TNF-alpha promoter polymorphism on the responsiveness to

inflix-imab in patients with rheumatoid arthritis Scand J Rheumatol

2004, 33:228-232.